Autonomy boundary for production changes
Your team is about to let an AI agent operate semi-autonomously on production: it can open PRs, and you're debating whether to also let it run database migrations and infra changes (Terraform apply) without a human approving each one. Make the call on where to set the autonomy boundary, and explain the principle that decides which actions an AI may take unattended versus which always require a human gate.
requires_human_gate(action: str, blast_radius: int, undo_cost: int) → bool["open_pr",1,1]outfalse["terraform_apply",2,3]outtrue["global_config_push",5,3]outtrueTreat the AI’s output as a draft to verify, not an answer to trust. Name the specific flaw and the input that triggers it, say how you’d catch it (tests, edge cases, reading critically), and how you’d re-prompt or decompose to get it right.
Vibe & agentic: describe the solution in plain language (or narrate it) and the coach grades your approach.